Food From Air

The startup that wanted to turn CO2 into protein.

Author

Nicolás Cerdeira
June 12, 2025

Hey — It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 5 most important things:

This issue is brought to you by Delve.

Let’s get into it.

How a missing SOC 2 killed $500k in ARR.AD

This one will sound familiar: it starts with a signed enterprise intent letter… and ends with an awkward “we’ll circle back next fiscal year.”

When you’re selling into big companies, the sales and compliance process can make or break your startup — and you usually don’t realize how painful it gets until you’re already in the thick of it. Here’s the body count:

  • Day 1: Enterprise deal - vendor security review arrives—80+ controls, 300 questions.

  • Day 30: Platform only at 40%, CTO still screenshot‑hunting; sales team sweating quota.

  • Day 60: Deal tabled, revenue evaporates, morale shot.

Lesson learned: Speed kills—or saves. Delve users finish the entire SOC 2 process in 15 hours, not 15 weeks, because:

  • Incredibly simplified workflow built for founders.

  • AI features to make the process lightning fast.

  • Delve managed audit done in 10 days or less.

Outcome:

  • Lovable closed SOC 2 in days.

  • Bland closed $500kk in ARR.

  • Sully got 8 frameworks under their belt.

Ready to script a different ending to your SOC 2? 

This Week In Startups

🔗 Resources

Sam Altman’s latest essay: The Gentle Singularity

What Is An AI-Native Company?

The internet killed general-purpose products. AI will bring them back.

📰 News

The Browser Company launches its AI-first browser, Dia.

Google’s AI search features are killing traffic to publishers.

Apple lets developers tap into its offline AI models.

💸 Fundraising

Sam Altman-backed Coco Robotics raises $80M.

Meter, a networking start-up, raises $170M.

Gecko Robotics raises $125 million.

Space Computing Startup Aethero Raises $8.4M.

Fail(St)ory

A New Kind of Food

A few days ago, Arkeon Biotechnologies filed for insolvency.

If you’ve never heard of it, Arkeon was one of those startups that sounded like magic. They turned carbon emissions into food—literally. They built a system where microbes ate industrial CO₂ and spit out amino acids, the building blocks of protein. No farmland, no animals, no crops. 

Just food made out of thin air.

What Was Arkeon:

Arkeon launched in 2021 in Vienna. Their goal was to completely reinvent how we produce protein.

They used a process called gas fermentation that involves feeding industrial CO₂ emissions to a special type of microbe—archaea. These microorganisms are ancient, resilient, and remarkably efficient at converting gas into usable nutrients.

Using this process, Arkeon developed a system that could produce all 20 essential amino acids in one step—no agriculture required. Just carbon and microbes.

It sounded like sci-fi. And honestly, it kind of was.

Arkeon’s big selling point wasn’t just environmental. Their ingredients were carbon-negative, fully vegan, GMO-free, and customizable for everything from plant-based food to supplements and even cultivated meat. They described their ingredients as “clean-label, functional proteins”—which means they were trying to appeal to both health-conscious consumers and large-scale food producers.

The Numbers:

  • 📅 Founded in 2021.

  • 💰 Raised $13M+ in venture capital.

  • 🏭 Opened a pilot facility in 2023.

  • 🧬 Used archaea microbes to produce all 20 amino acids from CO₂.

Reasons for Failure: 

  • Tough Market Timing: Arkeon raised most of its funding before 2023—back when capital was cheap and alternative proteins were still the next big thing. Since then, interest has shifted sharply. Funding for alt-protein startups dropped 44% in 2023 and another 27% in 2024. Investors who once chased climate tech are now writing checks for AI startups instead.

  • Underfunded Compared to Competitors: Similar startups like Air Protein and Solar Foods raised over $100M and $76M. Arkeon only pulled in $13M total. In biotech, that’s the difference between running a pilot and building a factory. They were never financially equipped to compete.

  • Regulatory Roadblock: Arkeon’s CEO mentioned “navigating regulatory landscapes” as a key lesson learned. And rightly so. In the EU, using archaea in food counts as a Novel Food, which means clearing 18–36 months of safety reviews with EFSA. That’s a long dead zone—no sales, no validation, no traction. Arkeon was racing regulators, not the market, and regulators aren’t in a hurry.

  • Difficulty Scaling: Producing a tiny amount of protein in a lab is one thing. Producing tons of it at industrial scale is another. Arkeon’s system depended on converting industrial CO₂ into food, but that also meant managing complex logistics, regulatory hurdles, and uncertain economics. Their tech worked—but scaling it didn’t.

Why It Matters: 

  • The case shows how fragile timing is for deep-tech startups. A great idea, launched at the wrong moment, might still fail.

  • It also points to a bigger truth: solving the protein problem will require more than just plant burgers. Traditional agriculture—even the plant-based kind—has real limits. Technologies like Arkeon’s might still be part of the answer. But they need time, funding, and a lot of operational discipline to get there.

  • Regulators don’t move fast. If your business depends on them, you need a long runway.

Trend

More Advanced Reasoning

Last week, we got two big updates from the AI world — and both point to the same thing: reasoning models are becoming the norm.

OpenAI launched a new model called o3-pro, which they’re calling their smartest one yet. At the same time, Mistral, the French startup behind many of the top open-source models, released Magistral, the first European reasoning model.

These two announcements show how reasoning AIs — models that think step-by-step rather than guessing answers — are starting to dominate the field. They’re not just getting better. They’re also getting cheaper, faster, and more distributed. 

A few months ago, it was just OpenAI in this space. Now there’s Google, Anthropic, DeepSeek, and with Mistral joining in, it’s clear this is becoming the next big frontier in AI.

Let’s break down what happened.

Why It Matters:

  • Smarter AI: Reasoning models can solve more complex problems and give more accurate responses.

  • Cheaper Access: The price of using top-tier models just dropped drastically.

  • More Players: Europe finally joins the reasoning race with Mistral’s Magistral.

OpenAI’s o3-pro

Let’s start with OpenAI. They launched o3-pro, a model designed for deep work — the kind where accuracy matters more than speed.

On paper, it’s impressive. It beats Google’s Gemini 2.5 and Anthropic’s Claude 4 on benchmarks like AIME and GPQA Diamond — tests built to push models in math, science, and logic.

But what makes this even more impressive is the price.

Compared to o1-pro, o3-pro costs 87% less. That’s $20 per million input tokens and $80 for output. 

In addition, the base o3 model got even more affordable — just $2 for input and $8 for output. That pricing puts OpenAI on par with Google’s Gemini and makes it much more accessible for startups and developers.

Mistral’s Magistral

Meanwhile, over in France, Mistral launched its Magistral line — the first European-built reasoning models.

There are two versions. Magistral Small is open-source and free to download. Magistral Medium is more powerful and runs on Mistral’s own API and chatbot platform, Le Chat.

It doesn’t outperform the American giants. Not yet. On most benchmarks, it lags behind Gemini and Claude. But it does one thing really well: speed. Mistral claims it delivers answers 10x faster on Le Chat. That might not win academic tests, but it wins user experience.

Magistral also claims to have better support for more languages — not just English and French, but Arabic, Russian, Chinese. That alone makes it useful in markets most US models don’t optimize for.

And the fact that part of it is open source? That matters. OpenAI, Google, and Anthropic keep their best models locked down. Mistral’s betting on openness as a wedge. 

Reasoning as the Default

Here’s what’s changing: reasoning models are no longer just for math nerds and coders. They’re getting cheaper, faster, and more available — and that means they might soon be the default models everyone uses.

Up until now, these models were niche. You’d only reach for them when you needed something complex: solving an equation, debugging code, writing a research summary. They were expensive, slower, and locked behind APIs with premium pricing.

But that’s shifting fast.

OpenAI just slashed prices. Mistral made theirs open source. And suddenly, what used to be a high-end tool is looking more like a standard feature. If you can get better answers — more structured, more accurate — for the same price, why wouldn’t you?

Help Me Improve Failory

How Was Today's Newsletter?

If this issue was a startup, how would you rate it?

Login or Subscribe to participate in polls.

That's all of this edition.

Cheers,

Nico